/llm_qlora

Fine-tuning LLMs using QLoRA

Primary LanguageJupyter NotebookMIT LicenseMIT

Fine-tuning LLMs using QLoRA

Setup

First, make sure you are using python 3.8+. If you're using python 3.7, see the Troubleshooting section below.

pip install -r requirements.txt

Run training

python train.py <config_file>

For exmaple, to fine-tune Llama3-8B on the wizard_vicuna_70k_unfiltered dataset, run

python train.py configs/llama3_8b_chat_uncensored.yaml

Push model to HuggingFace Hub

Follow instructions here.

Models trained on HuggingFace Hub

Model name Config file URL
llama3_8b_chat_uncensored configs/llama3_8b_chat_uncensored.yaml https://huggingface.co/georgesung/llama3_8b_chat_uncensored
llama2_7b_openorca_35k configs/llama2_7b_openorca_35k.yaml https://huggingface.co/georgesung/llama2_7b_openorca_35k
llama2_7b_chat_uncensored configs/llama2_7b_chat_uncensored.yaml https://huggingface.co/georgesung/llama2_7b_chat_uncensored
open_llama_7b_qlora_uncensored configs/open_llama_7b_qlora_uncensored.yaml https://huggingface.co/georgesung/llama2_7b_openorca_35k

Inference

Simple sanity check:

python inference.py

For notebooks with example inference results, see inference.ipynb and this Colab notebook.

Blog post

Blog post describing the process of QLoRA fine tuning: https://georgesung.github.io/ai/qlora-ift/

Converting to GGUF and quantizing the model

Download and build llama.cpp, and follow the instructions on their README to convert the model to GGUF and quantize to desired specs.

Tip: If llama.cpp gives an error saying the number of tokens is different between the model and tokenizer.json, it could be because we added a pad token (e.g. for training Llama). One work-around is to copy the original tokenizer.json from the base model (you can find the base model in huggingface cache at ~/.cache/huggingface/) to the new model's location, but make sure to back-up your tokenizer.json!

Tip: Llama3 uses BPE tokenizer, make sure to specify --vocab-type bpe when converting to GGUF

Troubleshooting

Issues with python 3.7

If you're using python 3.7, you will install transformers 4.30.x, since transformers >=4.31.0 no longer supports python 3.7. If you then install the latest version of peft, the GPU memory consumption will be higher than usual. The work-around is to use an older version of peft to go along with the older transformers version you installed. Update your requirements.txt as follows:

transformers==4.30.2
git+https://github.com/huggingface/peft.git@86290e9660d24ef0d0cedcf57710da249dd1f2f4

Of course, make sure to remove the original lines with transformers and peft, and run pip install -r requirements.txt